부산대학교 도서관

Recognize text in natural scenes is a challenging task. We proposed an attention-based deep neural network architecture for scene text recognition, which integrates feature extraction, feature attention, feature labeling and transcription into a unified framework. The primary advantages of the proposed model are: (1) it is an end-to-end model, does not require any segmentation of the input image. Convolutional neural network (CNN) is used as encoder to extract features, recurrent neural network (RNN) is used as decoder based on its characteristics of predict sequence, which composed a encoder-decoder architecture; (2) Soft Attention mechanism is introduced in, to further extract features in the input image, and allowing for end-to-end training within a standard back propagation framework; (3) Experiments are performed on several challenging scene text datasets, including IIIT5K, Street View Text, ICDAR2003 and ICDAR2013. Results of the experiments show that the proposed model is comparable or better than other models, which demonstrate the superiority of the proposed algorithm.